Construct a sharable GPU farm for Data Scientist. Layne Peng DellEMC OCTO TRIGr

Transcription

1 Construct a sharable GPU farm for Data Scientist Layne Peng OCTO TRIGr

2 Agenda Introduction Problems of consuming GPU GPU-as-a Service (project ) How to manage GPU in farm? How to schedule jobs to GPU? How to integrate with cloud platform? Next step and contacts 2 of 15

3 Technology Research Innovation Group (TRIGr) Innovation Advance Research Proof of Concept User Feedback Agile Roadmap 3 of 15

4 About me Wechat: Layne Peng, Principal Technologist at OCTO, experienced in cloud computing and SDDC areas, responsible for cloud computing related initiatives since joint in 2011, 10+ patents and a big data and cloud computing book author. 4 of 15

5 What is GPU & GPGPU? Graphical Processing Unit Designed for Parallel Processing Latency: Increase Clock Speed Throughput: Increase Concurrent Task Execution General-Purpose GPU Why: Scientific Computing and Graphical processing => Matrix & Vector Prereuisite: key feature supported in GPU, such as float point computing Easy to use: CUDA => Theano, Tensorflow and etc. Widely used in machine learning and deep learning area 5 of 15

6 Problems of consuming GPU today Not widely deployed High cost Low availability when sharing Hard to monitoring in fine-grained ü Remote GPU as a Service ü Fine-grained resource control ü Sharable ü Fine-grained resource monitor and tracing Offering GPU in cloud makes it more complex: How to abstract GPU resource? Pass-through solution today Performance issues caused by network Resource discovery and provisioning 6 of 15

7 Project GPU-as-a Service Remote GPU Access Consume remote GPU resource transparently Based on ueue model and various network fabric GPU Sharing N:1 model: multiple user consume same GPU accelerator Fine-grained control and management GPU Chaining 1:N model: multiple GPU hosts chaining as one to serve a large size to 1 application; Dynamical GPU number fit to application Smart Scheduling GPU pooling, discovery and provision Fine-grained resource monitor and tracing Pluggable intelligent scheduling algorithms: o Heterogeneous resource o Network tradeoffs 7 of 15

8 Overview of project Providing transparent remote GPU resource access: interception Intelligent selection of network fabric: TCP or RDMA Abstract the GPU resource to : Device from client s view, same with GPU Small management grained C GPU Library C Network (TCP, RDMA) GPU farm: GPU hosts management Sharable resource pool GPU Library Chaining features Intelligent scheduling Client Network Fabric C : CUDA Application 8 of 15

9 How to manage GPU in farm? Manage GPU farm: Consul based Basic cluster features: Add new nodes Remove nodes Detect failure Expel nodes and resurrect nodes Key takeaways: Abstract the hardware resource in a enough level Learn the experience from MSA Manager Manager a. Reuest to add to GPU farm Note: Master: a executable binary based on Consul with adding features; It can also deploy to the node with GPU; : is an extension of Consul with checking scripts and special configuration. Manager b. Reuest/Force to leave GPU farm c. Failure detected and expel failed node Cluster link Action 9 of 15

10 How to select GPU hosts to chain? GPU can be chained and exposed as one large machine to application GPU hosts selected for chaining by: Capacity 10 of 15 Queue length (crowdedness) Network radius Hops Bandwidth Key takeaways: Network is the first citizen of considering which GPU serve to application in remote case and chaining servers 3. Offer s to client Allocation API Global Scheduler XaaS_ 2. Chaining GPU from hosts selected to serve as one big machine has nine Network Radius 1. Select the nodes according to capacity and network radius C GPU Library Selected hosts

11 How to schedule jobs to GPU hosts? Client Application Two-level scheduling: 1 st level: find GPU hosts from farm GPU Library 2 nd level: find device for CUDA reuests Global smart scheduling: Global status monitor and trace Network topology and analysis Queue-based priority Pluggable scheduling strategy Network Strategy Host Dispatcher Q-Length Strategy H-Res Strategy Intelligent Analysis Distributed Global State Store Key takeaways: Think resources scheduling in a global view Frontend Smart Dispatcher Tasks Queue 11 of 15

12 Leverage ML to solve heterogeneous scheduling? Heterogeneous scheduling: Heterogeneous hardware resource Heterogeneous network fabric NP-hard problem Greedy Algorithm Or? Pluggable scheduler design: Abstract of scheduler API Machine learning jobs has work load pattern Metadata matters for scheduling Use machine learning to solve the scheduling problem of machine learning Status: on-going Training Phase Learn 3. Execute the Job Cluster Manager Model training 1. Job reuests Sync nodes attributes & status 2. Allocation Strategies Inference Phase Model trained Externel_Scheduler Colocation_Scheduler Allocation API Scheduler API Global Scheduler XaaS_ 12 of 15 : Job : Allocation Strategy

13 How to integrate with cloud platform? Offering allocation strategies as a service: High-level abstraction of allocation/scheduling functions Non-invasion design, easy to integrate with existed scheduler using in Cloud Native Platform 3. Execute the Job Cluster Manager 1. Job reuests Allocation API Global Scheduler XaaS_ Access GPU transparently from applications deployed in Cloud Native Platform Ease of integrating to Cloud Native Platform: Mesos/Marathon Kubernetes Cloud Foundry (demo in World 2017) 4. Transparently consume GPU as a service 2. Allocation Strategies Sync nodes attributes & status 13 of 15 : Job : Allocation Strategy

14 Example: Cloud Foundry Integration 1. Deploy machine learning task/application with CLI As a friendly machine learning interface at to project Receive jobs from users Manage cluster Start the containers for jobs Cloud Foundry offering Machine learning platform doesn t need to consider how environment configure and hardware accelerations Run my app with GPU! I don t care how! Reach out: Victor Fong from Dojo Data Scientist CLI GUI 2. Provision GPU service for application according to XaaS_ allocation strategies, and bind to applications Allocation API Global Scheduler XaaS_ Cloud Foundry Cloud Diego Brain Loggreator Marketplace Diego Cell Diego Cell Router Users 3. Users consume the task or application. GPU Farm 14 of 15

15 Next step and contacts Intelligent scheduling Schedule according to application metadata Model and calculate network distance Introduce machine learning to solve heterogeneous resource scheduling Keep improve performance Intelligent data movement and loading Enterprise features such as snapshot and etc. XaaS More hardware resource as a service, such as FPGAaaS Global-wide maximized performance Reach out the team Jack Harwood: jack.harwood@dell.com (distinguished engineer in ) Frank Zhao: junping.zhao@dell.com (consultant engineer, leads data path) Layne Peng: layne.peng@dell.com 15 of 15

16